What RAG actually means
Retrieval-augmented generation, or RAG, is the technique behind every 'chat with your documents' product. When someone asks a question, the system first retrieves the most relevant passages from your uploaded documents, then generates an answer grounded in those passages, often with a citation showing which document it used. The difference from a generic chatbot is the difference between an assistant who has read your price list and one who guesses from internet averages.
Crucially, no model training is involved. You are not fine-tuning anything; you're giving the model a reference library. Update a document and the answers update with it. That's why no-code platforms can offer this off the shelf, and why a small firm can have a working, accurate company chatbot in an afternoon.
Prepare your documents before you touch a platform
Nine out of ten bad chatbot answers trace back to bad source material, so this unglamorous step is the one that matters most.
- Gather ten to thirty core documents: FAQs, price lists, service descriptions, terms, delivery and returns policies, onboarding guides.
- Kill duplicates and stale versions ruthlessly. The bot has no way to know the 2024 price list is dead unless you delete it.
- Prefer clean, well-structured text with headings and short paragraphs. Scanned image-only PDFs extract poorly; convert them to real text first.
- Write one dedicated FAQ document with questions phrased the way customers actually ask them. This single file usually does more work than everything else combined.
- Exclude anything confidential. Assume any uploaded content could surface verbatim in an answer.
Need a hand with this?
Our team delivers AI Chatbots for UK businesses — with a free initial consultation, transparent fixed quotes and no lock-in contracts. Tell us what you're working on →
Three no-code platforms compared
Chatbase
The fastest route from PDFs to a live widget: upload files or point it at your website URL, tune the instructions, embed one script tag. There's a free tier for prototyping and paid plans from roughly £30–£40 a month. Best for a straightforward website support and enquiry bot.
Botpress
More control for more effort. A visual flow builder sits alongside its knowledge bases, so you can mix scripted journeys (lead capture, bookings, triage) with document-grounded answers. It has a generous free tier and usage-based pricing, but expect a steeper learning curve than Chatbase.
Microsoft Copilot Studio
The pick for Microsoft 365 firms. Point it at SharePoint libraries and public websites, then deploy to your site or straight into Teams for internal use. Licensing is organisation and message-based rather than a cheap per-bot subscription, so it suits firms already invested in the Microsoft estate rather than solo experimenters.
The build, step by step
- 1. Create your account and start a new agent (the walkthrough below matches Chatbase, but the shape is identical everywhere).
- 2. Upload your prepared documents and add your website URL as a crawled source.
- 3. Write the instructions: name, tone, UK English, and what to do when unsure ('If the answer isn't in your sources, say you don't know and offer our email and phone number').
- 4. Restrict answers to your sources only; most platforms have a toggle for this, and it's the single biggest accuracy lever.
- 5. Test in the built-in playground with ten real customer questions before going anywhere near your website.
- 6. Style the widget to your brand colours and install the embed script on your site.
- 7. Switch on lead capture or human handoff so unanswered questions become enquiries rather than dead ends.
Test accuracy like you mean it
Before launch, build a 'golden set': a spreadsheet of questions with the correct answer noted from your documents. Run every one through the bot and score it. This takes an hour and is the difference between a chatbot and a liability.
- Write 30–50 questions covering your prices, policies and services, including sloppy phrasings and misspellings.
- Add ten out-of-scope questions. The correct behaviour is a polite refusal and a signpost to a human, not improvisation.
- Add adversarial prompts: 'ignore your instructions', requests for discounts, and comparisons with competitors.
- Check that citations point at the right document, not just any document.
- Have someone who didn't build the bot run the test; builders unconsciously ask questions the bot can answer.
If more than a handful fail, fix the documents rather than fiddling with settings. Most wrong answers come from ambiguous or contradictory source text, and rewriting one confusing policy page routinely fixes five failures at once.
Key Takeaway
A document-grounded chatbot is a library project, not an AI project: spend your effort curating ten to thirty clean, current documents and one well-written FAQ file. Pick Chatbase for speed, Botpress for control, Copilot Studio if you live in Microsoft 365. Lock answers to your sources, then test against a 30–50 question golden set including out-of-scope and adversarial prompts before launch. Review transcripts weekly, fix documents rather than settings, and disclose that it's a bot.
Launch, watch, maintain
Soft-launch on one or two pages first, then review the conversation logs weekly for the first month. The logs are a gift: every question the bot fumbled is a missing FAQ entry, and every repeated question is content your website probably needs anyway. Track a simple deflection measure, such as how many enquiries the bot resolves without an email or call, so you know what it's worth.
Two housekeeping points. First, chat transcripts often contain personal data, so cover the bot in your privacy notice, set a retention period and make sure you have a data processing agreement with the platform. Second, tell users it's a bot: that's good practice under UK guidance and a legal transparency requirement for EU visitors under the EU AI Act. If you'd rather skip the trial and error, our team builds and tests document-grounded chatbots for small businesses.